Re: [Starpu-devel] StarPU 1.2.3 - Serious bug in OpenCL

[Alexander Grund <Alexander.Grund@tu-dresden.de>]

Hi Samuel,

thank you for the quick reply and fixes. Unfortunately it is not fixed 
yet and still deadlocks.
My test: Run axpy example on an Intel iGPU via OpenCL.

First a comment about the clFlush after the read/write: IMO the clFlush 
is to soon. The general pattern is:
enqueue read/write
if<OK> then
  if<synchronous> then wait for event
  else store/return event

The only place where a clFlush is required is the last one: <OK> && 
<asynchronous> as one does not need a flush if there was an error 
(obviously) or we clWaitForEvents (implicit flush 
https://www.khronos.org/registry/OpenCL/sdk/1.0/docs/man/xhtml/clFlush.html)

And second about the persisting bug: You added the clFlush after the 
kernel is enqueued. But after that the event that will be waited for is 
enqueued but not flushed. So instead of the clFlush on line ~961 you'd 
better add it on line ~1051 after the marker was enqueued and the error 
checked. This is shorter as the queue has already been requested and 
better as it is only done for the asynchronous case where it is required.

I tested that change and it works on my platform.

Alex


Am 16.02.2018 um 15:49 schrieb Samuel Thibault:
> Hello,
>
> Alexander Grund, on ven. 16 févr. 2018 10:05:23 +0100, wrote:
> > However the action is never flushed to the device, which requires 'clFlush'.
> > Hence the events status is CL_QUEUED and never changes.
> Oh, right, we are missing a flush here.  I have now added it (and ported
> to 1.2 and 1.3 branches).
>
> > Furthermore a few of the links on the downloads section are broken.
> > Basically 5 of the 6 "nightly snapshot" links (linux leads to HTTP 404 and
> That is due to our very recent migration to git, which changed the
> layout a bit, and some parts haven't been refreshed apparently.
> It's "branches/" which is spurious in the URL
>
> > windows to a jenkins login)
> Oh, indeed. Nathalie, do you think we can make that part of jenkins
> public?  (I'd say we could as well make our whole jenkins results public
> actually)
>
> > and the "Coverage" link of 1.2.3 (typo in the link: it has an extra
> > "2")
> Oops, indeed!
>
> (unfortunately I don't seem to have access to these, so we'll have to
> wait for Nathalie to come back from vacation)
>
> Samuel

--
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Technische Universitaet Dresden
Zentrum für Informationsdienste und Hochleistungsrechnen (ZIH)
Verteiltes und Datenintensives Rechnen (VDR)
Dipl.-Ing. Alexander Grund
Chemnitzer Str. 46b, Raum 243 01062 Dresden
Tel.: +49 351 463-35982
E-Mail: alexander.grund@tu-dresden.de
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~